OpenAI said on Thursday, June 4 that it is rolling out a new ChatGPT memory architecture, dubbed Dreaming V3, to US Plus and Pro subscribers first, with Free and Go users in more countries to follow over the next several weeks. The company said "recent improvements reduced the compute required to serve dreaming to Free users by approximately 5x," finally making always-on memory practical for the unpaid tier.
From Saved Memories to Always-On Dreaming
Saved memories — the explicit "remember this" cues OpenAI introduced in April 2024 — get replaced as the primary memory layer by Dreaming V3's background synthesis of past chats. The first dreaming release in April 2025 was a supplement; the new version makes the background process the standalone base. ChatGPT now keeps what The Decoder describes as a "coherent prose profile" of the user, sorted into categories such as work, hobbies and travel, and rewrites stale entries automatically (turning an upcoming Singapore trip into a finished one once the dates pass).
Benchmarks Are OpenAI's Own
OpenAI reported that Dreaming V3 hits 82.8% on fact retrieval (up from 67.9% in 2025 and 41.5% in 2024), 71.3% on preference following (up from 31.4%) and 75.1% on freshness (up from 52.2%). The figures come from internal tests with no rival-model comparison disclosed. Paid users also get twice the memory capacity and a new summary page for reviewing what ChatGPT may use to personalize responses, building on the GPT-5.5 Instant sources panel shipped in May.
Privacy Researchers Push Back
A February arXiv paper accepted at the ACM Web Conference 2026 analyzed 2,050 ChatGPT memory entries from 80 users and found that 96% were created unilaterally by the system rather than by explicit user instruction, with 28% containing GDPR-defined personal data and 52% holding psychological insights. The same automated-memory pattern has appeared in wrongful-death and product-safety suits against OpenAI in 2026, and users who prefer the old system can revert at Settings > Memory > Saved memories.
EU AI Act Clock
OpenAI's broader country expansion lands just months before transparency obligations under the EU AI Act take effect in August 2026 — a deadline that has shaped many of this quarter's AI rollouts, from Anthropic's Project Glasswing expansion to Microsoft's MAI-Code-1 push and Apple's upcoming WWDC 2026 Siri overhaul.
Reporting based on coverage from Implicator.ai, OpenAI's release notes, The Decoder and Engadget.